The proliferation of video cameras, such as those deployed in security surveillance setups, has drastically increased the amount of collected video data. Because human review of camera footage is tiresome and expensive, some processing techniques have been developed that enable the pre-selection of areas-of-interest within the feed for which processing is to be performed. However, such processing methods require that the user identify regions of interest before recording, a constraint that is often impractical for security deployments in which a location of events-of-interest within the frame is often hard to predict.